Human-Robot Emotional Interaction Model Based on Reinforcement Learning

Abstract

Emotional interaction is pivotal in fostering collaboration between humans and robots. Current paradigm in humanrobot interaction predominantly addresses physical aspects, overlooking the nuanced emotional dynamics inherent in the interaction. In this paper, an Emotional Interaction Reinforcement Learning (EI-RL) framework is proposed to address the gap. The framework is adept at generating emotionally responsive behaviors grounded in dynamic emotion polarity. Furthermore, the emotional reward function integrates the principle of interpersonal similarity from social psychology. To operationalize the framework, we utilize a continuous dialog corpus sourced from the MELD dataset, enabling the computation of an emotion transfer matrix essential for constructing an emotionally intelligent reinforcement learning environment. Empirical validation of the proposed framework was conducted through human-robot interaction experiments employing the Pepper robot. Our experimental results underscore the efficacy of the proposed approach in augmenting the quality of human-robot interactions.